Robust Visual Perception
Abstract
Over the last few years, deep neural networks have established themselves as the state-of-the-art technology for building computer-based visual perception systems, often surpassing human performance on a wide range of tasks. Despite impressive progress, however, it is becoming increasingly evident that current deep learning models lack robustness guarantees that are needed to deploy them safely. For example, it is possible to evade state-of-the-art surveillance cameras by wearing carefully-crafted color patches that confuse the underlying computer vision models. To address this problem, our goal is to develop novel methods for specifying and assessing robustness of deep neural models. Concretely, we target three key parts of the existing development process: (i) specification, by providing a rich fragment of semantic robustness specifications for computer vision models, (ii) assessment of the model robustness, which is automated and does not need to be manually tailored and supervised by an expert, and (iii) training models not only robust with respect to the specified robustness properties but also achieve state-of-the-art accuracy. We expect these results to boost the adoption of current techniques for adversarial testing and training by making them significantly more useful and practical.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 22, 2020
- Source ID
- W911NF2010317
Entities
People
- Petar Tsankov
Organizations
- Army Contracting Command
- LatticeFlow
- United States Army